VQGAN + CLIP

VQGAN + CLIP

Turns a text phrase into an image you can watch form from noise, and saves the whole process as a video. Suits artists, students and developers curious about the classic text-to-image method.

Made by multimodalart on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space brings back the classic VQGAN plus CLIP method of making images from text. You type a phrase and an image slowly takes shape from random noise, steered by how well it matches your words. Every step of that process is kept, so you also get a video of the image forming.

Who it suits: It suits people who like generative art, students learning how early text-to-image systems worked, and developers who want to read a compact implementation.

Industry | Who uses it | How they use it | Example
Digital art | Artists | Make abstract images from a phrase and keep the process video | Share the clip of an image forming
Education | Students and teachers | Watch how a text-guided image is built step by step | Show a class how noise turns into a picture
Research | ML hobbyists | Compare four guidance models on the same phrase | Try ViT-B/16 against SigLIP2 on one prompt
Software | Developers | Read the generation and video code | Look at how frames are saved without extra GPU time

You can skip it if you want photo-real results from modern image models, since this is a recreation of an older method.

Where to use it:

• A web browser: open the Space page on Hugging Face
• Windows, macOS or Linux: run a copy with Python
• A computer with a capable GPU for larger images, since memory limits the frame size

Use case | Who | What they get back
Generate from a phrase | Artists and hobbyists | An image shaped by your words, with a live preview
Save the process | Creators | An MP4 of every step, about 10 seconds for 250 steps at the default rate
Set a custom size | Designers | Frames from 128 to 1280 pixels, with square, wide and tall presets
Start from your own picture | Experimenters | An image seeded from an optional starting image

How to set it up: The online page needs nothing. A local copy needs Python and the packages in requirements.txt.

• A web browser for the online demo
• Python for a local copy
• The packages listed in requirements.txt

1. To use it online, open the Space page on Hugging Face.
2. To run a copy, install the Hugging Face hub package.
3. Download the Space into a local folder.
4. Move into the folder and install Gradio 6.20.0 and the listed packages.
5. Start the app.

pip install -U huggingface_hub
hf download multimodalart/vqgan-clip-video --repo-type space --local-dir vqgan-clip-video
cd vqgan-clip-video
pip install gradio==6.20.0
pip install -r requirements.txt
python app.py

How to use it: Run the app from its folder, or open the Space online. Type a phrase, pick one of the four guidance models, and start the run. The preview refreshes every 20 steps, and when the run ends you get the full video of the descent.

python app.py

Example: You type a short phrase describing a scene and use the default settings of 512 by 512 pixels and 32 cutouts. You watch the preview change as noise turns into a picture, then download the MP4 of the whole run.

For newcomers and experts: If you are new to this, type one short phrase and leave every setting alone. Watching the picture appear from static is the point, so keep the video.

If you know the field, you pick among four models that score how well the image fits your words: three CLIP variants and one SigLIP2. All four load at start. Each prefers a precompiled package and falls back to eager bf16 when that package is missing or mismatched. Frames are copied through pinned buffers into a raw file, then encoded to MP4 on the CPU, so the video costs no GPU time. With the compile option off, width and height move in steps of 16, matching the codebook's downsampling.

You type a phrase and pick a guidance model
An image starts as noise or your seed image
Each step moves it toward your words
Every step's frame is saved
The frames are joined into an MP4

Good to know: Memory, not the slider, limits the frame size, because the decoder and the cutout batch share one budget. At 32 cutouts the largest good sizes depend on the guidance model.

Title: Largest good frame width at 32 cutouts
CLIP ViT-B/32: 1280 pixels
CLIP ViT-B/16: 1024 pixels
CLIP ViT-L/14: 768 pixels
SigLIP2: 768 pixels

• License: MIT
• Cost to run: none named
• Data it reaches: the phrase and optional starting image you give it
• Not for: sizes beyond what the guidance model's memory allows, where the status line says so

On Hugging Face, 13 November 2026

• Likes: 6
• Built with: gradio
• License: mit

Links

• Space: https://huggingface.co/spaces/multimodalart/vqgan-clip-video
• Files: https://huggingface.co/spaces/multimodalart/vqgan-clip-video/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.